Register in variationist linguistics

Benedikt SzmrecsanyiKU Leuven

Benedikt Szmrecsanyi, Professor of Linguistics in the Quantitative Lexicol-ogy and Variational Linguistics research group at the Katholieke Universiteit(KU) Leuven, writes this article exploring the connections between registerand variationist linguistics. He is involved with various large-scale researchprojects in areas such as probabilistic grammar, variationist sociolinguisticresearch, linguistic complexity, and dialectology/dialectometry. Szmrec-sanyi’s books include Grammatical Variation in British English Dialects: AStudy in Corpus-based Dialectometry (2013, Cambridge) and AggregatingDialectology, Typology, and Register Analysis: Linguistic Variation in Text andSpeech (Szmrecsanyi & Wälchli 2014, Mouton de Gruyter). He is currently aprincipal investigator on a major grant-funded research project titled ‘Theregister-specificity of probabilistic grammatical knowledge in English andDutch’, a project aimed at exploring the question of whether register differ-ences lead to differences in the processes of making linguistic choices. Insharp contrast to the status quo in variationist linguistics, where register isoften ignored entirely, much of Szmrecsanyi’s variationist research treatsregister as a variable of primary importance. The findings from these studieshave led Benedikt Szmrecsanyi to state that “we need more empirical/varia-tionist work to explore the differences that register makes” (Szmrecsanyi2017:696).

Keywords: variationist sociolinguistics, corpus linguistics, probabilisticgrammar

1. How is register conceptualized in variationist linguistics?

Variationist linguistics is a discipline in the field of variation studies that inves-tigates variation between “alternate ways of saying ‘the same’ thing” (Labov1972: 188). In this spirit, variationist linguists carefully account for competing vari-ants and draw on quantitative methodologies to model the conditioning factorsthat regulate the way language users choose between semantically and function-ally equivalent variants.

https://doi.org/10.1075/rs.18006.szmRegister Studies 1:1 (2019), pp. 76–99. issn 2542-9477 | e‑issn 2542-9485© John Benjamins Publishing Company

Variationist linguistics and traditional register research thus adopt different,not always compatible perspectives on language variation and its intersectionwith what we may call the register-genre-style triad. According to customarydefinitions (see, e.g., Biber & Conrad 2012:8), analyzing registers means strictlyspeaking investigating the functional relationship(s) between a set of linguisticfeatures and the situational context. Variationist linguistics, however, is aboutvariation between different ways of accomplishing the same function (Labov1972: 188). Therefore, functional variation does by definition not come under theremit of variationist linguistics because attention is typically restricted to func-tionally equivalent forms (see also Biber & Conrad 2004:265 on this point). Thisis why variationist linguistics is largely agnostic about functional relationships.

Against this backdrop, it is maybe not surprising that there is quite a bit ofterminological variance regarding the register-genre-style triad in the variation-ist literature: some variationists use the term ‘register’ (e.g. Gries 2015), some usethe term ‘genre’ (e.g. Grafmiller 2014), and variationist sociolinguists in particu-lar focus on ‘style’ (e.g. Eckert & Rickford 2001), which is often conceptualizedas the amount of attention paid to one’s speech. In the corpus-based variationistliterature, the concept that most researchers really have in mind when they usethe terms ‘register’ and ‘genre’ comes, a tad confusingly, close to what Biber andConrad (2012:2) refer to as ‘style’: aesthetic preferences that cause speakers andwriters to favor particular variants in particular situations (and this also happensto be the conceptualization that underlies the case study to be presented later inthis article). In actual research practice, of course, many corpus-based variation-ist linguists tend to rely on the register/genre distinctions built into the design ofexisting publicly available corpora (consider, e.g., the Brown corpus family withits 15-fold register/genre categorization – see Table 1 below) without losing muchsleep over functional relationships. Sometimes, register is treated as an outrightnuisance factor that needs to be controlled for in the analysis so as not to distortresults. I am not aware of variationist work where the goal is the discovery of rele-vant register distinctions.

A second major difference in research philosophy between register researchand variationist linguistics is that register analysts often aggregate over multiplefeatures to characterize registers and register differences, the rationale being thatit is hard to find individual features that comprehensively characterize registers(Biber & Conrad 2012:9). By contrast, orthodox variationist analysis eschewsaggregation and proceeds in what Nerbonne (2009: 176) calls a “single-feature-based” mode: researchers explore individual linguistic variables, one variable at atime (typically, one variable per research paper). It is certainly true that the fieldis moving towards multi-variable designs (see, e.g., Guy 2013, Guy & Hinskens2016), but the one-variable-at-a-time approach is still the customary one.

Register in variationist linguistics 77

A third point of difference is that while in traditional register research, thetypical unit of observation is individual texts, variationist analysis is focused –as well shall see below – on individual linguistic choices. The task in variationistanalysis is therefore to link register differences not to text frequencies but to lin-guistic choice-making.

In summary, then, in variationist linguistics register is typically conceptual-ized as stylistic variation in aesthetic preferences. Register is thought of as oneof the language-external factors (beside, e.g., real time, geographic provenance,etc.) that regulates variation of individual linguistic variables. Register is specifi-cally analyzed in terms of how it influences linguistic choices between function-ally equivalent variants.

2. How does register relate to the research goals within variationistlinguistics?

The central research goal in variationist linguistics is to understand how peoplechoose between different ways of saying the same thing. Specifically, variationistlinguists study the constraints that regulate linguistic choice-making. Constraintscan be language-internal (e.g., the phonetic environment in choice contexts) orlanguage-external (e.g., demographic factors, geography, etc.). Register, then, isone of the language-external factors that needs to be included in good modelsof how people choose between linguistic variants, for the sake of better under-standing how language-internal and language-external constraints shape linguis-tic variation. I exemplify by considering the well-known alternation in the gram-mar of English between the prepositional dative construction, as in (1a), and theditransitive dative construction, as in (1b):

(1) a. I’ve sent a message to him via a couple of different channels […](Corpus of Global Web-based English GB G)

b. I sent him a message saying I am simply going out with friends […](Corpus of Global Web-based English AU G)

Röthlisberger, Grafmiller, and Szmrecsanyi (2017) study this syntactic alternationbased on corpus material sampling multiple geographic varieties of English anda range of registers. To understand how variation is constrained in the mate-rial at hand, the study fits a logistic regression model predicting dative choicebased on well-known language-internal constraints (e.g., animacy of the recip-ient, length of the constituents) and two language-external factors, variety ofEnglish (e.g., British English versus Canadian English) and register (e.g., spo-ken-informal texts versus spoken-formal texts). Analysis shows that variety is

78 Benedikt Szmrecsanyi

a more important factor than register, but at the same the two factors interactsignificantly, which means that stylistic norms vary across varieties of English.Accordingly, in the Röthlisberger et al. (2017) study, register is quite central tothe research goal of variationist linguistics: understanding how variation works.

With that being said, it is fair to say that register is a language-externalfactor that has received less attention in the variationist literature than other fac-tors, such as social factors. Note in this connection, however, that the field isnot entirely homogeneous: in Szmrecsanyi (2017), I draw a distinction between(Labovian) variationist sociolinguistics (also known as the ‘Language Variationand Change paradigm’), and corpus-based variationist linguistics. Among otherthings, variationist sociolinguists are more interested in social determinants ofvariation than corpus-based variationist linguists, who typically focus more onlanguage-internal constraints on variation. But also, because corpus-based varia-tionist linguists typically rely on large, publicly available corpora that often sam-ple multiple text types, register variation is a topic that is comparatively moreimportant in corpus-based variationist linguistics. Sometimes register is concep-tualized as a key factor of substantial interest (e.g. in Szmrecsanyi 2006, whoinvestigates among other things how persistence/priming effects differ acrossregisters), and sometimes register is modeled as a factor that may not be of pri-mary interest but that needs nonetheless attention so as to not distort results andfor the sake of accounting for variation (this, for instance, is the spirit of theRöthlisberger et al. 2017 study discussed above). But in either case, corpus-basedvariationist linguists do care about register variation. By contrast, it seems fair tosay variationist sociolinguists of the Labovian persuasion tend to be less inter-ested in register. It is true that style and style-shifting have received ample atten-tion in variationist sociolinguistics (see e.g., Bell 1984, Rickford & McNair-Knox1994). But, variationist sociolinguists rarely consider, e.g., non-spoken texts, andtend to be especially interested in vernacular speech as manifested in sociolin-guistic interviews: the credo is, in a nutshell, that “variation in language is mostreadily observed in the vernacular of everyday life” (Tagliamonte 2012: 2; seeD’Arcy & Tagliamonte 2015 for critical discussion). There are two reasons whyvernacular speech has a special status in theorizing in variationist sociolinguis-tics: For one thing, vernacular speech is considered “the style in which the min-imum attention is given to the monitoring of speech” (Labov 1972:208), so thevernacular is where variation is thought to be at its best (unlike, as the reason-ing goes, written language, which is more “governed by prescription”; D’Arcy& Tagliamonte 2015:255). Second, more practically speaking, the bulk of workin variationist sociolinguistics deals with phonetic variables, where written textsand, to some extent, formal speech are irrelevant.

Register in variationist linguistics 79

What have we learned about register and register variation through varia-tionist linguistics? As far as the variationist sociolinguistic literature is concerned,there is plenty of evidence that style shifting across spoken styles, which can beseen as a form of register variation, is an excellent diagnostic of the social mean-ing of variation. As Rickford and Eckert (2001: 1) put it, “style is a pivotal constructin the study of sociolinguistic variation”, which is why sociolinguistic interviewsare designed to elicit a range of speech styles. Historically speaking, variation-ist sociolinguistic interest in style goes back to Labov (1966). As indicated above,Labov essentially defined style as the attention paid to one’s speech. Style shift-ing, then, indicates the perceived prestige of variants. For example, in the famousdepartment store study Labov (1966) found that many department store employ-ees are more likely to pronounce postvocalic /r/ in careful, emphatic pronuncia-tion, which demonstrates that rhoticity is prestigious. The generalization is thatcareful speech increases the rate of higher-prestige variants, while casual speechincreases the rate of lower-prestige variants. At the same time, style-shifting inter-acts with age and sex: normally, there is more style-shifting in female speech thanin male speech (Eckert 2000: 195). It should be added that more nuanced perspec-tives on style have been developed over the years, especially in what is known asthird-wave sociolinguistics (Eckert 2018), but it is not clear that this work comesunder the remit of variationist linguistics as defined here.

As to the corpus-based variationist literature, register has been shown timeand again to be an important factor regulating variation. Recent studies that havefound a significant main effect of register under multivariate control, or substan-tial random effects, include the following:

– Heylen (2005) investigates constituent order variation in varieties of Germanand finds that the difference between the spoken and written medium has asignificant impact on variant choice;

– Grondelaers, Speelman, and Geeraerts (2008) model the Dutch postverbal er(‘there’) retention versus omission alternation and show that register (UseNetdiscourse versus popular newspapers versus quality newspapers) is a signifi-cant factor in Belgian Dutch;

– Levshina (2011: Chapter 6), in her study of the Dutch doen versus laten alter-nation, reports some significant register effects, with doen being particularlyunlikely to be used in conversations and disfavored in web-based Dutch(Usenet discourse) in comparison to Dutch from newspapers;

– Lohmann (2011) analyzes the help versus help to alternation in English andfinds that including genre distinctions significantly improves model accuracy;

– Wolk, Bresnan, Rosenbach, and Szmrecsanyi (2013) study the historicaldevelopment of, among other things, the English dative alternation in A

80 Benedikt Szmrecsanyi

Representative Corpus of Historical English Registers (ARCHER) and reportthat factoring in register differences improves model accuracy;

– Pijpops and Van De Velde (2014) find significant register effects (chat versusemail versus quality newspaper versus tabloid) in some of their models of theDutch partitive genitive alternation;

– Gries (2015) investigates the English particle placement alternation and findsthat what he calls “sub-registers” (110) (e.g. private versus public dialogue,scripted versus unscripted monologue, and so on) matter for predicting par-ticle placement choices;

– Rosemeyer and Enrique-Arias (2016) model the conditioning of variation inthe expression of possession in Old Spanish and report that in their Bible cor-pus, register variation (lyrical, narrative) has a significant effect on variantchoices;

– Szmrecsanyi et al. (2016) investigate ternary genitive variation in the lateModern English period and find that including register information (newsversus science versus letters) as a random effect improves model accuracy;

– Heller (2017) is concerned with the genitive alternation across a range of post-colonial varieties of English and reports that the distinction between writtenand spoken registers has a significant effect on variant choice;

– Grafmiller and Szmrecsanyi (in press) study the particle placement alterna-tion across varieties of English and report that according to conditional ran-dom forest analysis, register variation (written formal versus written informalversus spoken formal versus spoken informal) often has a substantial impacton variant choices.

The cumulative weight of evidence from research on style-shifting and corpus-based variationist linguistics thus suggests that register regularly1 affects the rela-tive frequency with which speakers and writers select particular variants. Specif-ically, register and style differences can be utilized as a diagnostic of prestigedifferentials, and in statistical models of grammatical variation based on corpusdata, register distinctions often improve model quality.

1. Failures to obtain significant register effects are not systematically reported in the literature,which is why it is hard to be more precise.

Register in variationist linguistics 81

3. What are the major methodological approaches that are used toanalyze or account for register in variationist linguistics?

To account for register in variationist linguistics, analysts apply the variationistmethod (see, e.g., Labov 1969, Sankoff 1988) and include register as (one of) thelanguage-external constraints on the variation phenomenon under study. Cacoul-los and Walker (2009:326–327) concisely define the gist of the variationist methodas follows:

[the variationist method] seeks to discover patterns of usage in the relative fre-quency of co-occurrence of linguistic forms and elements of the linguistic con-text. The interpretative component of the variationist method lies in identifyingsimilar discourse functions of different constructions […] We account for theselection of variants to fulfill a particular discourse function by exhaustivelyextracting each instance of that function in discourse and applying quantitativetechniques to determine the influences of contextual factors on the choice ofform.

In practice, conducting a variationist analysis of register effects consists of the fol-lowing steps:

1. Selection of the variable: The analyst picks one (and traditionally, only one)variation phenomenon (also known as a linguistic variable or – in the realmof grammar – an alternation). If the analyst has a particular interest in howregister shapes variation patterns, she will want to select a variation phenom-enon that we have reason to believe is particularly sensitive to register differ-ences and/or style-shifting.

2. Circumscription of the variable context: Guided by the Principle of Account-ability (“any variable form […] should be reported with the proportion ofcases in which the form did occur in the relevant environment, compared tothe total number of cases in which it might have occurred”, Labov 1969: 738,n. 20), the analyst catalogs all variant forms and properly circumscribes thevariable context to ensure that attention is restricted to only those contexts inwhich language users truly have a choice between all competing variants.

3. Retrieval and annotation: Based on the circumscription of the variable con-text in step 2, the analysts next turns to production/corpus data and identifiesand extracts all relevant variant forms in the material. Subsequently, the ana-lyst annotates each variant form for language-internal and language-externalconstraints on variation. To identify the set of potentially relevant language-internal constraints, analysts typically survey the literature and/or rely onintuitions. The language-external constraints are typically determined on a

82 Benedikt Szmrecsanyi

by-text or by-transcript basis (i.e., register, real time, demographic character-istics of the speaker/writer, and so on).

4. Analysis: The richly annotated dataset generated in step 3 is then analyzed sta-tistically to determine the conditioning of variation using multivariate analy-sis methods such as logistic regression analysis or conditional random forestanalysis (see, e.g., Tagliamonte & Baayen 2012 for an accessible introduction).

5. Interpretation: What do the results generated in step 4 reveal about the inter-action between register and linguistic variation?

In variationist analysis, the unit of observation is thus individual linguisticchoices, and not, e.g., texts (see Biber, Egbert, Gray, Oppliger, & Szmrecsanyi 2016for discussion). Relevant research questions about register include the following:How do particular registers influence the odds that people select particular vari-ants? How powerful a predictor is register vis-à-vis other language-external con-straints? Also, in terms of explanatory power, how does register fare vis-à-vislanguage-internal constraints?

4. What does a typical register study look like in variationist linguistics?

4.1 Selection of the variable

As an empirical case study, we will now investigate variation between relativizerwhich, as in (2a), and relativizer that, as in (2b):

(2) a. The largest hurdle the Republicans would have to face is a state law whichsays that before making a first race, one of two alternative courses must be

(Brown text A01)takenb. Fulton legislators work with city officials to pass enabling legislation that

will permit the establishment of a fair and equitable pension plan for city(Brown text A01)employees

Which and that in contexts such as (2) qualify as different ways of saying the samething. The research question to be addressed in this case study is the following:How important a factor is register vis-à-vis other constraints in shaping varia-tion between the explicit relativizers which and that? We know that which-thatvariation is subject to a number of language-internal probabilistic constraints.We also know that written English – particularly American English – is driftingtowards increased usage of that because of a process that Hinrichs, Szmrecsanyi,and Bohmann (2015:806) call “institutionally backed colloquialization”. Finally,we know that which is the incoming form that has a bookish feel to it, while thatis the more colloquial variant that is widespread in vernacular language (Biber,

Register in variationist linguistics 83

Johansson, Leech, Conrad, & Finegan 1999: 610; Tagliamonte, Smith & Lawrence2005). In summary, then, which-that variation is richly constrained by language-internal and language-external constraints, which makes this alternation an inter-esting phenomenon to study from a variationist perspective.

4.2 Circumscription of the variable context

I adopt the circumscription of the variable context used in Hinrichs et al. (2015).Attention is thus restricted to:

– finite relative clauses introduced by which and that, ignoring, e.g., participialrelative clauses of the type the man standing at the bar;

– restrictive relative clauses, because standard English non-restrictive relativeclauses (as in [3]) only allow which. The criterion for restrictiveness is theabsence of a comma preceding the relativizer, which in standard written Eng-lish is a sufficiently reliable diagnostic;

– subject relative clauses (i.e., clauses where the relativizer acts as subject),because restrictive non-subject relative clauses also permit zero as a relativizer(see [4]);

– relative clauses with inanimate antecedents, as animate antecedents tend totrigger the relativizers who/whom/whose (as in [5]).

Lastly, the study ignores oblique relatives with pied-piping (as in [6]) becausethese categorically take which.

(3) The jury further said in term-end presentments that the City Executive Com-mittee, which had over-all charge of the election, deserves the praise and

(Brown text A01)thanks of the City of Atlanta

(4) […] that what we were asserting to be bad was precisely the suffering ∅ we(Brown text J52)thought had occurred back there […]

(5) Barber, who is in his 13th year as a legislator, said there are some members ofour congressional delegation in Washington who would like to see it (the res-

(Brown text A01)olution) passed.

(6) […] and concentrate its constructive efforts on eliminating in other parts ofLatin America the social conditions on which totalitarian nationalism feeds

(Brown text A04)

84 Benedikt Szmrecsanyi

4.3 Retrieval and annotation

We will re-analyze2 a subset of the relativizer dataset analyzed in Grafmiller, Szm-recsanyi, and Hinrichs (2016), which, in turn, largely overlaps with the relativizerdataset analyzed in Hinrichs et al. (2015). The dataset is publicly available as supple-mentary materials to Grafmiller et al. (2016) <https://doi.org/10.1515/cllt-2016-0015>.Variable relativizer tokens were extracted from Brown, Frown, LOB, and F-LOB(see Hinrichs, Smith, & Waibel 2010 and references therein). Each corpus con-tains roughly 1 million words of written text of standard American (Brown, Frown)and British (LOB, F-LOB) English compiled in the early 1960s (Brown, LOB) andthe early 1990s (Frown, F-LOB). Each corpus consists of 500 2,000-word samplesrepresenting data from 15 distinct ‘categories’ (in Brown parlance), e.g. newspaperarticles, humor writing, academic writing, and various genres of fiction. These cate-gories can be grouped into four ‘genre groups’ (again, in Brown parlance): (1) press,(2) general prose, (3) learned, and (4) fiction.

Extraction of variant forms was in line with the description of the variablecontext detailed above (in addition to subject relative clauses, the dataset also cov-ers non-subject relative clauses, which however are not analyzed in the presentstudy). To extract instances of relative clauses with overt relativizers, the compilersmade extensive use of the Part-Of-Speech tagging available in the four corpora.For additional details on the extraction methods, see Hinrichs et al.(2015: 815–816). In total, the subject relativizer dataset to be analyzed in the presentpaper spans N=4,400 which tokens and N=5,731 that tokens.

All tokens were annotated for a suite of constraints thought to influence thechoice of relativizer. In the present study, I consider a smaller set of constraintswhich, according to the analysis in Hinrichs et al. (2015) and Grafmiller et al.(2016), are particularly important:

– Preceding relativizer (precrel). Which relativizer was used last time thewriter had a choice? The predictor distinguishes the levels ‘that’, ‘which’, ‘zero’,and ‘none’ (for when the relativizer in question is the first one encountered ina corpus file). Consider (7), where the choice context the primes which enter ispreceded by the choice context The theorem which we prove.

(7) The theorem which we prove is more general than what we havedescribed since it works with the primary decomposition of the minimalpolynomial whether or not the primes which enter are all of first degree.

(Brown text J18)

2. The analyses were carried out using the R software environment for statistical computing<https://www.r-project.org/>. The R script is available at <https://osf.io/k4rfp/>.

Register in variationist linguistics 85

This predictor gauges the effect of priming or structural persistence (see, e.g.,Gries 2005; Szmrecsanyi 2005): language users tend to recycle recently usedgrammatical variants in upcoming discourse. A univariate frequency analy-sis indicates that some 78.5% of that relativizers are preceded by another that,while 74.0% of which relativizers are preceded by another which.

– Part of Speech of antecedent (antpos). The annotation distinguishes twolevels, ‘noun’ vs. ‘other’, to gauge the effect of antecedent pronominality inde-pendently from definiteness, and to distinguish lexically specific antecedentsfrom ‘empty’ ones, as in (8).

(8) all that happens is that the better qualified teacher declines to gamble twoor three years of his life on the chance that conditions at the Catholic insti-

(Brown text A35)tution will be as good as those elsewhere.

The analysis in Hinrichs et al. (2015: Table 4) indicates that nominalantecedents favor which, and in the dataset subject to analysis here whichshows larger occurrence rates when the antecedent is nominal (44.5%) thanwhen it is non-nominal (35.3%). The reverse is true for that (55.5% versus64.7%).

– Length of antecedent in words (antln). This measure gauges the complexityof the noun phrase modified by the relative clause. Consider (9), where theantecedent (the escheat law) has a length of three words.

(9) He told the committee the measure would merely provide means ofenforcing the escheat law which has been on the books “since Texas was a

(Brown text A02)republic”.

The analysis in Hinrichs et al. (2015: Table 4) suggests that increasingantecedent length disfavors that and favors which. In the dataset under analy-sis here, antecedent of that-relative clauses have a mean length of 3.24 words,while which-antecedents have a mean length of 3.54 words.

– Length of relative clause in words (rcln). This measure approximates thecomplexity of the clause introduced by the relativizer. Consider (9), where therelative clause (which has been on the books “since Texas was a republic”) has alength of 11 orthographic words. The analysis in Hinrichs et al. (2015: Table 4)suggests that increasing relative clause length disfavors that and favors which.In the dataset under analysis here, the mean length of that-relative clauses is8.80 words, while which-relative clauses have a mean length of 10.22 words.

86 Benedikt Szmrecsanyi

– Passive-active ratio (passiveactiveratio). This predictor measures theproportion of passive constructions (as in [10]) over active lexical verbs in agiven corpus text.

(10) (Brown text A02)His contention was denied by several bankers […].

The analysis in Hinrichs et al. (2015: Table 4) suggests that increased usage ofthe passive voice correlates with reduced that-usage. Likewise, in the datasetunder analysis here, the mean proportion of passive constructions is 0.40 forthat-relatives, while it is 0.70 for which-relatives.

– Time (time). This is a binary predictor (‘1960s’ vs. ‘1990s’) indicating thetime period when the text was sampled. The analysis in Hinrichs et al. (2015:Table 4) shows that texts from the 1990s favor usage of that, compared to textsfrom the 1960s. In the dataset under analysis, the proportion of that-relativesto which-relatives was 46.7%: 53.3% in the 1960s, but 66.1%: 33.9% in the 1990s.

– Variety (variety). This is a binary predictor indicating the variety of writtenStandard English (‘American English’ vs. ‘British English’) from which the textwas sampled. The analysis in Hinrichs et al. (2015: Table 4) suggests that allother things being equal, American English strongly favors that compared toBritish English. In the dataset under analysis, the proportion of that-relativesto which-relatives is 71.5%: 28.5% in American English, but only 41.0%: 59.0%in British English.

– Genre group (genregroup). This predictor distinguishes the followinggenre groups: (1) press, (2) general prose, (3) learned, and (4) fiction. In thedataset under analysis, the proportion of that-relatives to which-relatives is72.5%: 27.5% in fiction, 54.5%: 45.5% in general prose, 43.4%: 56.6% in learnedwriting, and 61.1%: 38.9% in press writing.

– Category (category). Relativizer proportions in the 15 categories covered inthe Brown family are displayed in Table 1.

– Corpus file (corpusfileid). This logs the particular corpus text in which arelativizer occurs, for the sake of modeling author idiosyncracies as a by-sub-ject random effect in regression analysis.

Register in variationist linguistics 87

Table 1. Variant rates by Brown categorythat which Total

A_Reportage 393 286 67957.9% 42.1% 6.7%

B_Editorial 304 214 51858.7% 41.3% 5.1%

C_Review 264 113 37770.0% 30.0% 3.7%

D_Religion 272 268 54050.4% 49.6% 5.3%

E_Skills 500 244 74467.2% 32.8% 7.3%

F_Popularlore 624 370 99462.8% 37.2% 9.8%

G_BellesLettres 901 829 173052.1% 47.9% 17.1%

H_Miscellaneous 226 399 62536.2% 63.8% 6.2%

J_Science 895 1165 206043.4% 56.6% 20.3%

K_GeneralFiction 286 127 41369.2% 30.8% 4.1%

L_Mystery 253 78 33176.4% 23.6% 3.3%

M_ScienceFiction 77 29 10672.6% 27.4% 1.0%

N_Adventure 383 116 49976.8% 23.2% 4.9%

P_Romance 251 98 34971.9% 28.1% 3.4%

R_Humor 102 64 16661.4% 38.6% 1.6%

Column Total 5731 4400 10131

4.4 Analysis

The distributional analyses provided in the foregoing discussion suggest that weare dealing with substantial variation between relativizer which and that, condi-tioned by both language internal-and language-external probabilistic constraints.The task before us is to model this variation statistically to determine the overallimportance, effect direction, and effect size of constraints on relativizer variationunder multivariate control. Multivariate control is essential to make sure that, e.g.,

88 Benedikt Szmrecsanyi

the high rate of which (63.8%) in the Miscellaneous category (see Table 1) is nottrivially due to the fact that relative clauses also happen to be longest (10.8 wordson average, versus a global mean of 9.4 words) in the Miscellaneous category. Mul-tivariate analysis techniques can take care of such confounds.

To evaluate the overall importance of the constraints, I begin by submittingthe dataset to conditional random forest (CRF) analysis as implemented in thecforest() function in R’s party package (Hothorn, Hornik & Zeileis 2006; Strobl,Boulesteix, Kneib, Augustin, & Zeilis 2008; Strobl, Boulesteix, Zeileis, & Hothorn2007). I skip a discussion of the technicalities and instead refer the reader to thevery accessible introduction in Tagliamonte and Baayen (2012). What is especiallyappealing from the point of view of the present study is that CRF can be usedto straightforwardly rank constraints according to their explanatory importance.This is the purpose of the dot plot in Figure 1, which ranks constraints on rela-tivizer variation according to their importance.

Figure 1. Conditional random forest permutation variable importance measures forconstraints on that versus which variation. C=0.90

According to CRF analysis, the most important constraint on relativizer varia-tion is precrel, the language-internal predictor that checks which relativizer wasused last time the writer had a choice. Strong persistence/priming effects in gram-matical variation have been reported before in the literature (e.g., Scherre & Naro1991; Weiner & Labov 1983), but it is noteworthy that the nature of the precedingrelativizer (precrel) is the single most important predictor of relativizer choice.The next three predictors in the ranking are all language-external in nature: vari-ety (British English versus American English), time (1960s versus 1990s), andcategory (which distinguishes between the 15 registers displayed in Table 1). In

Register in variationist linguistics 89

other words, regional differences are the most important language-external con-straint, while register differences are least important – though it should be stressedthat registers differences are still more central than most of the language-internalpredictors in the model except precrel. For exploratory reasons, I also includedthe rather coarse-grained predictor genregroup (press versus general prose ver-sus learned versus fiction) in the analysis, but Figure 1 shows that its importance islimited. This appears to suggest that we really do need the full 15-fold distinctionin the predictor category to responsibly model register variation. The predictorpassiveactiveratio scores a bit higher than relative clause length (rcln). Lastly,according to CRF analysis properties of the antecedent (antpos and antln) arefairly insubstantial.

Having thus determined the overall importance of the constraints understudy, I now turn to mixed-effects binary logistic regression analysis (Pinheiro& Bates 2000) as implemented in R’s lme4 package (Bates, Mächler, Bolker, &Walker 2015) to learn more about effect directions and effect strengths. Logisticregression analysis has been the workhorse analysis technique in variationist lin-guistics for decades – the Varbrul program (Cedergren & Sankoff 1974) is essen-tially an implementation of logistic regression. Logistic regression modeling is theclosest a corpus analyst can come to conducting a controlled experiment; the tech-nique models the combined contribution of all the factors considered in the analy-sis, systematically testing the probabilistic effect of each factor while holding theother factors in the model constant. In addition to so-called ‘fixed effects’, whichare classically estimated predictors that assess the reliability of the effect of repeat-able characteristics, mixed-effects modeling also includes ‘random effects’ to cap-ture variation dependent on open-ended, potentially hierarchical and unbalancedgroups (see Wolk et al. 2013: 399–400 for more discussion). As fixed effects, Ientered all of the language-internal factors described above plus variety andtime; as random effects, I included intercept adjustments for category (to gaugeregister differences) and corpusfileid (to approximate author idiosyncrasies as aso-called by-subject effects). Note here that the analyst has some leeway in decid-ing which factor(s) should be integrated into the fixed-effects structure, and whichinto the random-effects structure. In the case study at hand, it would certainlyhave been possible to model, e.g., variety as a random effect (after all, the British/American distinction is not necessarily repeatable, as the next study down theroad may include, e.g., Australian English), or possibly category as a fixed effect.But beside conceptual considerations (is a predictor repeatable or not?), morepragmatic considerations often play a role; regardless of the conceptual status ofthe predictor in question, the analyst may be justified in modeling a predictor as arandom effect, in order not to waste precious degrees of freedom (see, e.g., Zuur,Ieno, Walker, Saveliev, & Smith 2009: 106 for discussion). Because the predictor

90 Benedikt Szmrecsanyi

category has no fewer than 15 levels, modeling the predictor as a random effectis an elegant solution. The downside is that random effect modeling cannot deter-mine if differences between levels are statistically significant or not.

Coefficients associated with the fixed effects in the resulting model are dis-played in Table 2. (The models omits the predictor genregroup, which CRFanalysis has shown to be rather dispensable).

Table 2. Coefficients of fixed effects in mixed-effects binary logistic regression analysis ofvariation between that versus which

Coefficient (b)

(Intercept) −2.61 ***

precRel (default: none)

that −0.32 ***

which  0.37 ***

zero −0.01

antPOS (default: non-nominal)

Nominal  0.59 ***

antLN  0.06 ***

RCLn  0.05 ***

passiveActiveRatio  0.50 ***

variety (default: American English)

British English  2.04 ***

time (default: 1961)

1991 −1.3 ***

Note. Significance codes: 0 – ‘***’; 0.001 – ‘**’; 0.01 – ‘*’. Predicted odds are for which. C=0.93; condi-tion number (kappa) =7.8;% outcomes correctly predicted: 80.6% (baseline: 56.6%).

The column ‘Coefficient (b)’ displays regression coefficients. Negative coef-ficients disfavor the predicted outcome, which is usage of which; positive coef-ficients favor the predicted outcome. Let us go through the coefficients one byone. The reference level for precrel is ‘none’ (this happens when the relativizerunder analysis is the first in a given corpus text). Compared to this condition, apreceding that relativizer significantly disfavors usage of which; but a precedingwhich relativizer significantly favors re-use of which (a preceding zero relativizerdoes not have a significant effect on choice between which and that). Needlessto say, these effects are perfectly in line with the literature on priming. Likewiseconsistent with the literature (Hinrichs, Szmrecsanyi & Bohmann 2015), nominalantecedents (antpos), long antecedents (antln), and long relative clauses (rcln)

Register in variationist linguistics 91

Figure 2. Intercept adjustments for random effect CATEGORY. Positive adjustmentsfavor which, negative adjustments favor that

all favor choice of which. Also, increased usage of the passive voice (passive-activeratio) increases the odds for which (in line with Hinrichs, Szmrecsanyi,& Bohmann 2015). As to the language-external predictors, observe that in com-parison to American English, British English favors which, while texts written inthe 1990s disfavor which, compared to texts written in the 1960s. In other words,we see a real time drift towards that, as described by Hinrichs, Szmrecsanyi &Bohmann (2015).

Let us finally address the issue most germane to this article: What is the effectthat register distinctions have on relativizer choice? Figure 2 displays the interceptadjustments (i.e., adjustments to the base probability that the relativizer which ischosen, all other things being equal) that the model makes for individual levels ofthe random effect category. The upshot is that the register categories that moststrongly favor which are – in descending order of magnitude – Miscellaneous (H),Science (J), and Belles Lettres (G). The register categories that disfavor which andinstead favor that are Mystery and detective Fiction (L), Science Fiction (M), andAdventure and Western (N). To come back to an issue mentioned earlier, then,Figure 2 provides multivariate evidence that, e.g., the high rate of which in theMiscellaneous category (see Table 1) is indeed not trivially due to the fact that rel-ative clauses also happen to be longest in Miscellaneous: There really is a stylisticpreference for which in that category.

92 Benedikt Szmrecsanyi

4.5 Interpretation

As to language-internal constraints on which-that variation, the most importantfactor is structural persistence: There is a surprisingly strong tendency to re-usethe relativizer which was used last time there was a choice, in line with the lit-erature on priming and related phenomena (see Szmrecsanyi 2006 for extendeddiscussion). The effect of most of the other language-internal constraints underconsideration in the present study are consistent with Rohdenburg’s complex-ity principle (Rohdenburg 1996), which stipulates that more explicit grammati-cal variants are used in cognitively more complex environments. Note now thatwhich can be argued to be somewhat more explicit than that, thanks, among otherthings to, the fact that it contains more phonetic material than that ([wɪtʃ] or even[hwɪtʃ] versus [ðət]). And as we have seen in regression analysis, it is which thatis favored when the antecedent is long (and thus arguably complex) and whenthe relative clause is long (and thus arguably complex). The positive correlationbetween the predictor passiveactiveratio and usage of which suggests that theincreasing popularity of that is not primarily an outcome of the fact that 20th cen-tury style guides prescribe that in restrictive relative clauses (“Careful writers […]go which-hunting, remove the defining whiches and by so doing improve theirwork”; Strunk & White 1999: 59). Instead, the variation patterns seem to be regu-lated by formality (see Hinrichs, Szmrecsanyi, & Bohmann 2015 for discussion):which-users tend to consistently opt for formal linguistic options.

The hierarchy of importance (Figure 1) of the three language-external con-straints we studied is variety > time > register. Register differences are thus lessimportant than geographic contrasts or real time drifts, but on the other handregister differences are more important than most of the language-internal con-straints except for precrel (priming). Analysis of the effect directions shows thatthe real-time drift towards relative that is significant, all other things being equal,and that it is led by American English.

With regard to register differences, we see that writers have aesthetic prefer-ences such that informational registers (e.g., science texts) are particularly hos-pitable towards which, while imaginative registers (e.g., mystery and detective fic-tion) are particularly hospitable towards that. This distribution certainly supportsthe established view in the literature (e.g., Biber et al. 1999: 610; Tagliamonte,Smith, & Lawrence 2005) that that is the more vernacular variant, compared to‘bookish’ which. Now, as we saw in the introductory section, interpreting variation-ist findings in terms of functional relationships is problematic because the varia-tionist methodology stricto sensu requires analysts to restrict attention to function-ally equivalent constructions and forms. But that being said, one could reasonably

Register in variationist linguistics 93

argue that the function of that is to establish a less formal, more vernacular tone,while the function of which is to establish a more formal, ‘bookish’ tone.

5. What are the most promising areas of future register research invariationist linguistics?

The research record suggests that register/genre/style distinctions are importantingredients in variationist accounts of the relative frequencies with which speakersand writers use linguistic variants. In the case study reported in the previous sec-tion, register happens to be a slightly less important factor overall than otherlanguage-external predictors (real time and regional differences), but other alter-nations may be differentially sensitive to register distinctions – further researchinto this issue is clearly needed.

Also, variationist linguists are going to have to address deeper theoreticalquestions about the register-genre-style triad. It is possible, and indeed likely, thatthe way that register has been modeled in the bulk of the previous literature is sim-plistic, in that register is typically modeled merely as a main effect (as was donein the case study in this article). This implicitly assumes that, yes, particular regis-ters may favor usage of particular variants, but also that the way language-internalconstraints (e.g., length effects) regulate variation stays constant across registers.But, we do not know about the extent to which language users, when they have thechoice between different ways of saying the same thing, draw on different choice-making processes in different registers.

This issue is loaded theoretically but under-investigated empirically. In thevariationist sociolinguistic community, there is a sense that variation grammarsare stable across styles:

For the most part, stylistic variation is quantitatively simple, involving raising orlowering the selection frequency of socially sensitive variables without alteringother grammatical constraints on variant selection; indeed, it is commonlyassumed in VR [Variable Rule, BS] analyses that the grammar is unchanged in

(Guy 2005: 562; see also Labov 2010: 265)stylistic variation.

Guy’s claim may indeed be correct if attention is restricted to style-shifting in soci-olinguistic interviews. It is not clear, however that this quantitative simplicity eas-ily generalizes to ‘real’ register variation, e.g., variation across speech and writing(see Szmrecsanyi 2017 for discussion), and so, it is odd that corpus-based varia-tionist linguists have tended to ignore this issue. A notable exception to this negli-gence is Grafmiller (2014), who, based on corpora such as the Switchboard corpus

94 Benedikt Szmrecsanyi

and the Boston University Noun Phrase Corpus, investigates the English genitivealternation, as in (1d1).

(11) a. I have in this lecture sketched alternatives to the standard way of narratingphilosophy 's history as part of the development of Western culture […]

(COCA 2017 ACAD)b. He seems to have a partial understanding of its place in the history of phi-

(COCA 2017 ACAD)losophy […]

The genitive alternation is conditioned by a range of semantic, syntactic andphonological constraints. Grafmiller models these constraints across some six dif-ferent registers/genres (conversation, learned writing, non-fiction, general fiction,western fiction, press). His analysis uncovers significant interactions between reg-ister and the probabilistic weights that language-internal constraints on genitivevariation have. In model 1 (p.482), for example, no less than five of the nine lan-guage-internal constraints under study (possessor animacy, possessor givenness,possessor/possessum length, type-token ratio, possessor text frequency) turn outto have significantly different effect sizes across registers. This can be interpretedas evidence that grammar is not unchanged in register variation and that languageusers do indeed adjust their choice-making processes to the situational context.

If we reasonably assume that English genitive variation is not entirely atypical,massive register specificity à la Grafmiller (2014) raises important theoretical andmethodological questions about the nature and scope of grammatical variation,and about the interaction of this variation with socioculture. Among other things,if it is indeed the case that different registers come with differently sized con-straints, then Guy’s Grammatical Difference Hypothesis (Guy 2015), according towhich having different (or differently sized) constraints means having a differ-ent grammars, would arguably warrant us to conclude that language users havea number of different register-specific variable grammars, akin to situations ofdiglossia or bilingualism.

Work is underway in the variationist community to investigate this issue morefully. For example, researchers at the KU Leuven and at the University of Birm-ingham (Principal Investigators: Benedikt Szmrecsanyi, Jason Grafmiller, FreekVan de Velde) have embarked in 2018 on a project entitled “The register-speci-ficity of probabilistic grammatical knowledge in English and Dutch” (funded bythe Research Foundation Flanders under grant # G0D4618N). The project inves-tigates parallel grammatical alternations in English and Dutch. Based on corpusstudy and on rating task experiments, the project seeks to establish the extent towhich language users adjust their choice-making to the situational context: Whatare the relevant register-related dimensions of variability: formal versus informal(formality), or written versus spoken (medium)? Do languages such as English

Register in variationist linguistics 95

and Dutch differ in terms of the importance of probabilistic register differences?And, what are the probabilistic constraints that tend to have particularly variableprobabilistic effects across registers?

These are the type of questions where the theoretically responsible intersec-tion of register studies and variationist linguistics can substantially advance theo-rizing in usage-based linguistics.

Acknowledgements

Thanks go to Alexandra Engel, Laura Rosseel, and Freek Van de Velde for helpful comments onan earlier draft of this paper.

References

Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models usinglme4. Journal of Statistical Software, 67(1). 1–48. https://doi.org/10.18637/jss.v067.i01

Bell, A. (1984). Language style as audience design. Language in Society, 13(2), 145.https://doi.org/10.1017/S004740450001037X

Biber, D., & Conrad, S. (2004). Corpus-based comparisons of registers. In C. Coffin,A. Hewings, & K. O’Halloran (eds.), Applying English grammar: Functional and corpusapproaches (pp. 40–56). London: Hodder Arnold.

Biber, D., & Conrad, S. (2012). Register, genre, and style. Cambridge: Cambridge UniversityPress.

Biber, D., Egbert, J., Gray, B., Oppliger, R., & Szmrecsanyi, B. (2016). Variationist versustext-linguistic approaches to grammatical change in English: Nominal modifiers of headnouns. In M. Kytö & P. Pahta (Eds.), The Cambridge handbook of English historicallinguistics (pp. 351–375). Cambridge: Cambridge University Press.https://doi.org/10.1017/CBO9781139600231.022

Biber, D., Johansson, S., Leech, G., Conrad, S., & Finegan, E. (1999). Longman grammar ofspoken and written English. Harlow: Longman.

Cacoullos, R., & Walker, J. A. (2009). The present of the English future: Grammatical variationand collocations in discourse. Language, 85(2), 321–354. https://doi.org/10.1353/lan.0.0110

Cedergren, H., & Sankoff, D. (1974). Variable rules: Performance as a statistical reflection ofcompetence. Language, 50(2), 333. https://doi.org/10.2307/412441

D’Arcy, A., & Tagliamonte, S.A. (2015). Not always variable: Probing the vernacular grammar.Language Variation and Change, 27(3), 255–285. https://doi.org/10.1017/S0954394515000101

Eckert, P. (2000). Linguistic variation as social practice: The linguistic construction of identity inBelten High (Language in Society 27). Malden, MA: Blackwell.

Eckert, P. (2018). Meaning and linguistic variation: The third wave in sociolinguistics.Cambridge: Cambridge University Press.

Eckert, P., & Rickford, J.R. (2001). Style and sociolinguistic variation. Cambridge: CambridgeUniversity Press.

96 Benedikt Szmrecsanyi

Grafmiller, J. (2014). Variation in English genitives across modality and genres. EnglishLanguage and Linguistics, 18(3), 471–496. https://doi.org/10.1017/S1360674314000136

Grafmiller, J., & Szmrecsanyi, B. (in press). Mapping out particle placement in Englishesaround the world. A case study in comparative sociolinguistic analysis. LanguageVariation and Change.

Grafmiller, J., Szmrecsanyi, B., & Hinrichs, L. (2016). Restricting the restrictive relativizer.Corpus Linguistics and Linguistic Theory, 14(2), 309–355https://doi.org/10.1515/cllt‑2016‑0015. <https://www.degruyter.com/view/j/cllt.ahead-of-print/cllt-2016-0015/cllt-2016-0015.xml> (1 March, 2018).

Gries, S. T. (2005). Syntactic priming: A corpus-based approach. Journal of PsycholinguisticResearch, 34(4), 365–399. https://doi.org/10.1007/s10936‑005‑6139‑3

Gries, S. T. (2015). The most under-used statistical method in corpus linguistics: Multi-level(and mixed-effects) models. Corpora, 10(1), 95–125. https://doi.org/10.3366/cor.2015.0068

Grondelaers, S., Speelman, D., & Geeraerts, D. (2008). National variation in the use of er“there”. Regional and diachronic constraints on cognitive explanations. In G. Kristiansen,& R. Dirven, Cognitive sociolinguistics: Language variation, cultural models, social systems(pp. 153–204). Berlin: De Gruyter. https://doi.org/10.1515/9783110199154.2.153

Guy, G.R. (2005). Letters to language. Language, 81(3), 561–563.https://doi.org/10.1353/lan.2005.0132

Guy, G.R. (2013). The cognitive coherence of sociolects: How do speakers handle multiplesociolinguistic variables? Journal of Pragmatics, 52, 63–71.https://doi.org/10.1016/j.pragma.2012.12.019

Guy, G.R. (2015). Coherence, constraints and quantities. Talk given at NWAV44, Toronto.Guy, G.R., & Hinskens, F. (2016). Linguistic coherence: Systems, repertoires and speech

communities. Lingua, 172–173, 1–9. https://doi.org/10.1016/j.lingua.2016.01.001Heller, B. (2017). Stability and fluidity in syntactic variation world-wide: The genitive

alternation across varieties of English. Unpublished PhD dissertation, KU Leuven.https://doi.org/10.1177/0075424216685405

Heylen, K. (2005). Zur Abfolge (pro)nominaler Satzglieder im Deutschen. Eine korpusbasierteAnalyse der relativen Abfolge von nominalem Subjekt und pronominalem Objekt imMittelfeld. Unpublished PhD dissertation, KU Leuven.

Hinrichs, L., Smith, N., & Waibel, B. (2010). Manual of information for the part-of-speechtagged, post-edited “Brown” corpora. ICAME Journal, 34, 189–231.

Hinrichs, L., Szmrecsanyi, B., & Bohmann, A. (2015). Which-hunting and the Standard Englishrelative clause. Language, 91(4), 806–836. https://doi.org/10.1353/lan.2015.0062

Hothorn, T., Hornik, K., & Zeileis, A. (2006). Unbiased recursive partitioning: A conditionalinference framework. Journal of Computational and Graphical Statistics, 15(3), 651–674.https://doi.org/10.1198/106186006X133933

Labov, W. (1966). The Social Stratification of English in New York City. Washington DC: Centerfor Applied Linguistics.

Labov, W. (1969). Contraction, deletion, and inherent variability of the English copula.Language, 45, 715–762. https://doi.org/10.2307/412333

Labov, W. (1972). Sociolinguistic patterns. Philadelphia, PA: University of Philadelphia Press.Labov, W. (2010). Principles of linguistic change. Vol. 3: Cognitive and cultural factors (Language

in Society 39). Malden, MA: Wiley-Blackwell. https://doi.org/10.1002/9781444327496Levshina, N. (2011). Doe wat je niet laten kan [Do what you cannot let]: A usage-based analysis

of Dutch causative constructions. Unpublished PhD dissertation, KU Leuven.

Register in variationist linguistics 97

Lohmann, A. (2011). Help vs help to: A multifactorial, mixed-effects account of infinitivemarker omission. English Language and Linguistics, 15(3), 499–521.https://doi.org/10.1017/S1360674311000141

Nerbonne, J. (2009). Data-driven dialectology. Language and Linguistics Compass, 3(1),175–198. https://doi.org/10.1111/j.1749‑818X.2008.00114.x

Pijpops, D., & Van de Velde, F. (2014). A multivariate analysis of the partitive genitive in Dutch.Bringing quantitative data into a theoretical discussion. Corpus Linguistics and LinguisticTheory, 10, 1–30. https://doi.org/10.1515/cllt‑2013‑0027. <https://www.degruyter.com/view/j/cllt.ahead-of-print/cllt-2013-0027/cllt-2013-0027.xml> (14 February, 2018).

Pinheiro, J.C., & Bates, D.M. (2000). Mixed-effects models in S and S-{PLUS}. New York:Springer. https://doi.org/10.1007/978‑1‑4419‑0318‑1

Rickford, J.R., & Eckert, P. (2001). Introduction: John R. Rickford and Penelope Eckert. InP. Eckert & J.R. Rickford (Eds.), Style and Sociolinguistic Variation (pp. 1–18). Cambridge:Cambridge University Press. https://doi.org/10.1017/CBO9780511613258.001. <http://ebooks.cambridge.org/ref/id/CBO9780511613258A010> (31 December, 2017).

Rickford, J.R., & McNair-Knox, F. (1994). Addressee-and topic-influenced style shift: Aquantitative sociolinguistic study. In D. Biber & E. Finegan (Eds.), Perspectives on register:Situating register variation within sociolinguistics (pp. 235–276). Oxford: Oxford UniversityPress.

Rohdenburg, G. (1996). Cognitive complexity and increased grammatical explicitness inEnglish. Cognitive Linguistics, 7, 149–182. https://doi.org/10.1515/cogl.1996.7.2.149

Rosemeyer, M., & Enrique-Arias, A. (2016). A match made in heaven: Using parallel corporaand multinomial logistic regression to analyze the expression of possession in OldSpanish. Language Variation and Change, 28(3), 307–334.https://doi.org/10.1017/S0954394516000120

Röthlisberger, M., Grafmiller, J., & Szmrecsanyi, B. (2017). Cognitive indigenization effects inthe English dative alternation. Cognitive Linguistics, 28(4), 673–710.https://doi.org/10.1515/cog‑2016‑0051

Sankoff, D. (1988). Sociolinguistics and syntactic variation. In F. J. Newmeyer (Ed.), Linguistics:The Cambridge survey (pp. 140–161). Cambridge: Cambridge University Press.https://doi.org/10.1017/CBO9780511620577.009

Scherre, M., & Naro, A. (1991). Marking in discourse: “Birds of a feather.” Language Variationand Change, 3, 23–32. https://doi.org/10.1017/S0954394500000430

Strobl, C., Boulesteix, A., Kneib, T., Augustin, T. & Zeileis, A. (2008). Conditional variableimportance for random forests. BMC Bioinformatics, 9(1), 307.https://doi.org/10.1186/1471‑2105‑9‑307

Strobl, C., Boulesteix, A., Zeileis, A., & Hothorn, T. (2007). Bias in random forest variableimportance measures: Illustrations, sources and a solution. BMC Bioinformatics, 8(1), 25.https://doi.org/10.1186/1471‑2105‑8‑25

Strunk, W., & White, E.B. (1999). The elements of style, 4th ed. Longman.Szmrecsanyi, B. (2005). Language users as creatures of habit: A corpus-based analysis of

persistence in spoken English. Corpus Linguistics and Linguistic Theory, 1(1), 113–150.https://doi.org/10.1515/cllt.2005.1.1.113

Szmrecsanyi, B. (2006). Morphosyntactic persistence in spoken English: A corpus study at theintersection of variationist sociolinguistics, psycholinguistics, and discourse analysis. Berlin:Mouton de Gruyter. https://doi.org/10.1515/9783110197808

98 Benedikt Szmrecsanyi

Szmrecsanyi, B. (2013). Grammatical variation in British English dialects: A study incorpus-based dialectometry. Cambridge: Cambridge University Press.https://doi.org/10.1017/CBO9780511763380

Szmrecsanyi, B. (2017). Variationist sociolinguistics and corpus-based variationist linguistics:Overlap and cross-pollination potential. Canadian Journal of Linguistics/RevueCanadienne de Linguistique, 62(4), 1–17. https://doi.org/10.1017/cnj.2017.34

Szmrecsanyi, B., Biber, D., Egbert, J., & Franco, K. (2016). Toward more accountability:Modeling ternary genitive variation in Late Modern English. Language Variation andChange, 28(1), 1–29. https://doi.org/10.1017/S0954394515000198

Szmrecsanyi, B., & Wälchli, B. (Eds.). (2014). Aggregating dialectology, typology, and registeranalysis: Linguistic variation in text and speech. Berlin: Walter de Gruyter.https://doi.org/10.1515/9783110317558

Tagliamonte, S. (2012). Variationist sociolinguistics change, observation, interpretation. Malden,MA: Wiley-Blackwell. <http://public.eblib.com/EBLPublic/PublicView.do?ptiID=819316>(29 August, 2013).

Tagliamonte, S., Smith, J., & Lawrence, H. (2005). No taming the vernacular! Insights from therelatives in northern Britain. Language Variation and Change, 17(1), 75–112.https://doi.org/10.1017/S0954394505050040

Tagliamonte, S.A., & Baayen, R. H. (2012). Models, forests, and trees of York English: Was/werevariation as a case study for statistical practice. Language Variation and Change, 24(2),135–178. https://doi.org/10.1017/S0954394512000129

Weiner, J., & Labov, W. (1983). Constraints on the agentless passive. Journal of Linguistics, 19,29–58. https://doi.org/10.1017/S0022226700007441

Wolk, C., Bresnan, J., Rosenbach, A., & Szmrecsanyi, B. (2013). Dative and genitive variabilityin Late Modern English: Exploring cross-constructional variation and change.Diachronica, 30(3), 382–419. https://doi.org/10.1075/dia.30.3.04wol

Zuur, A. F., Ieno, E.N., Walker, N. J., Saveliev, A. A., & Smith, G. (2009). Mixed effects modelsand extensions in ecology with R. New York: Springer.https://doi.org/10.1007/978‑0‑387‑87458‑6

Address for correspondence

Benedikt SzmrecsanyiDepartment of LinguisticsKU LeuvenBlijde-Inkomststraat 21B-3000 LeuvenBelgiumbenszm@kuleuven.be

Register in variationist linguistics 99